An Innovative Approach of Textile Fabrics Identification from Mobile Images using Computer Vision based on Deep Transfer Learning Antonio Carlos da Silva Barros ‡§ , Elene Firmeza Ohata *§ , Suane Pires P. da Silva *§ , Jefferson Silva Almeida §† and Pedro Pedrosa Rebouc ¸as Filho *§ * Programa de P´ os-Graduac ¸˜ ao em Engenharia de Teleinform´ atica (PPGETI), Universidade Federal do Cear´ a, Fortaleza, Cear´ a, Brazil † Programa de P´ os-Graduac ¸˜ ao em Engenharia El´ etrica (PPGEE), Universidade Federal do Cear´ a, Fortaleza, Cear´ a, Brazil ‡ Universidade da Integrac ¸˜ ao Internacional da Lusofonia Afro-Brasileira (Unilab), Fortaleza, Cear´ a, Brazil § Laborat´ orio de Processamento de Imagens, Sinais e Computac ¸˜ ao Aplicada (LAPISCO), Instituto Federal do Cear´ a, Fortaleza, Cear´ a, Brazil Email: {carlosbarros, elene.ohata, suanepires, jeffersonsilva}@lapisco.ifce.edu.br, pedrosarf@ifce.edu.br Abstract—The identification of different textile fabrics is a task commonly learned in practice and, therefore, is considered a very strenuous and costly form of learning, causing annoyance to the individual who performs it. Based on this context, this paper proposes a new method for classifying textile fabrics, based on the development of a computer vision system using Convolutional Neural Network (CNN). CNN works as a feature extractor by incorporating the concept of Transfer Learning. Using Transfer Learning allows a pre-trained CNN model to be reused for a new problem. In order to highlight the high performance of CNN, an analysis is performed with feature extractors established in the literature. Parameters such as Accuracy, F1-Score, and processing time are considered to evaluate the efficiency of the proposed approach. For the classification were used Bayesian Classifier, Multi-layer Perceptron (MLP), k-Nearest Neighbor (kNN), Random Forest (RF), and Support Vector Machine (SVM). The results show that the best combination is the CNN architecture DenseNet201 with SVM (RBF), obtaining an accuracy of 94% and F1-Score of 94.2%. Index Terms—Textile Fabrics, Convolutional Neural Network, Transfer Learning, Computer Vision I. I NTRODUCTION The need for man to use a fabric that can cover his body goes back to the beginnings when there was a need for man to protect his body from cold or heat. The use and manufacture of fabrics became a basic need and later an item of luxury and social status. Clothing evolved with humanity and became a reflection of social, political, religious, and moral aspects of all stages experienced by human beings [1], [2]. Nowadays, because of the Industrial Revolution and sub- sequent scientific progress, few people need to know how to spin or weave, but they need to know how to judge the quality of the cloths made by the machines considering their duration. Therefore, the study of textiles becomes essential for all consumers, as well as the manufacture of the cloth, and the identification of the fiber assumes greater significance [1], [2]. There are people who learn to judge the quality of fabrics, through practice and experience, over time, but the method of ”learning by making mistakes” is costly and full of hassles for a human being [1], [2]. The computer vision area has solved several problems. It is a research field that has helped humankind to customize and automate different tasks. For instance, it can be used to aid in the textile industry tasks. In the literature, some works have applied this field of knowledge to the classification of textile fibers, classification of flat fabrics, detection of defects, or inspection. In [3], an automatic system for identification of fabric structure was developed, employing the principal component analysis (PCA) and fuzzy clustering. In [4], the authors used the Local Binary Patterns and Gray-Level Co- occurence Matrix, along with artificial neural networks, to detect fabric defects. In [5], the authors proposed an algorithm for detecting defects in fabrics, which was based on biological vision modeling. In [6], a method based on lattice segmen- tation and lattice templates was developed to automatically identifies the defects of fabric images. While in [7], the authors presented a method for detecting fabric defects based on autoencoder. In [8], the authors used a CNN to identify fabrics, but they used a dataset with 19,894 images. This article proposes an innovative approach to classify tex- tile fabrics. The approach consists of the use of CNN for fea- ture extraction, based on the definition of Transfer Learning. We evaluated the deep extractors with five classifiers. Because it is a complex method, vision-based classification relies on accurate calculations and fast processing times. Thus, to verify the performance of each classifier, two evaluation metrics were used: Accuracy (Acc) and F1-Score (F1S). Criteria such as extraction time and classification time were also measured. The results show that DenseNet201, DenseNet169, and DenseNet121 combined with SVM reached 94.35%, 93.34%, and 93.52%, respectively, in Accuracy, demonstrating to be 978-1-7281-6926-2/20/$31.00 ©2020 IEEE